Chad - trigger development
Analysis last updated 2022-01-03
For Chad an OCHA anticipatory action framework is being developed to anticipate drought. Such a framework consists of activities, financing, and a trigger. This document summarizes the work done by the Centre for Humanitarian Data (CHD), whose work focuses on the development of the trigger. The development of the trigger in Chad has been a collaborative process for which a preliminary trigger was proposed by a technical working group consisting of several in-country organizations. The validation and fine-tuning of this trigger was thereafter executed by CHD. This document is a summary, more elaborate analyses and code can be found on GitHub.
Context
Chad has several climatological zones. For this pilot the focus is on a part of the Sahelian zone, see the map below, as this area was identified as most vulnerable to drought. More specifically the ADMIN1 areas of Lac, Kanem, Barh-El-Gazel, Batha, and Wadi Fira are the focus. For the trigger it was decided to also only include this area of interest within the analyses, i.e. we look at the drought within the area and not at e.g. a possible decreased trade due to drought in other areas.
The focus of the trigger is on the rainy season. In Chad the rainy season is from June to September, with most rainfall in July and August. Below the historical median rainfall per month is shown. The average monthly rainfall over all the whole area and all years is 14mm in June, 61mm in July, 89mm in August, and 29mm in September.
Preliminary trigger
The proposed preliminary trigger is as follows:
Proposed preliminary trigger
Trigger validation
Based on the prelimanry trigger we did a validation and fine-tuning. This process consisted of four steps: 1) compiling a list of historical socio-economic drought years 2) extract the observed values of the trigger indicators at the end of the rainy season and see how these correlate with the list of historical drought years 3) extract the date of the same trigger indicators but that is forecasted or available during the rainy season and see how these correlate with the observed end of season data. 4) if no strong correlation in 2) or 3) test correlations with the same indicators but at different thresholds.
Historical socio-economic drought
With anticipatory action we want to anticipate droughts that cause an increase in humanitarian needs. Such droughts are often referred to as socio-economic droughts. To validate the trigger we can thus look at the past and see how well the moments the trigger is reached correspond with historical years we would have wanted to be triggering. To do so we thus need a list of historical socio-economic droughts. This kind of impact data is not directly available. It is scattered across sources and often has no detailed information on the temporal and spatial scale.
Nevertheless, we attempted to compile a list of worst drought years. We solely focus on the year, so not specific months, and look at Chad in general so not at specific areas as this data is often not available. The compiled list is by no means fully validated and shouldn’t be seen as a ground-truth dataset. We therefore use the list for part of the analyses but don’t base all decisions on it. The list was compiled in discussion with CERF colleagues and based on several sources.
The first source we can look at are the historical CERF allocations, which are available since 2006. In Chad drought-related allocations were in 2010-01, 2011-12, and 2018-05. By analyzing the linked documents we can conclude that these allocations were in response to poor rainy seasons during 2009, 2011, and 2017.
Another source is FewsNet which maps the food insecurity situation since 2009. In their reports the causes of food insecurity are often mentioned and thus we can use these reports to look for a mention of meteorological drought. We scanned these reports manually for the years there were CERF allocations. The 2010 report indicates increased food insecurity caused by erratic patterns during the 2009 rainy season. In 2011 a shorter than usual rainy season is mentioned. 2017 mentions localized dry spells causing bad pastoral conditions. This thus matches the allocations.
Another source is the official drought declarations. These are known to be in 2005/2006, 2009/2010, 2016/2017 but we don’t have any source of these. It thus remains unclear to which rainy seasons they relate. [What to do with them?! Especially the 2005 one we havent seen anywhere else]
Lastly, we can inspect other sources of literature such as the African Risk Capacity 2016/17 operations plan, the 2017 multi-risk contigency plan and the worldbank statistics. [we have the documents of the first two but couldn’t find them online]
ARC reports 1968 to 1973, 1983, 1984, 2004, and 2009 as drought years.
The multi-risk contigency plan reports the years 1969, 1981, 1993, 1997, 2001, 2009, and 2012. The world bank reports the years 1993, 1997, 2001, 2009, 2012, 2017. These three sources all don’t provide more information on the cause of the drought and it remains unclear whether these years refer to the rainy season or the lean season.
All in all, we can see that the different sources show similarities but also differences. Moreover, with the reports of FewsNet and CERF we see that socio-economic drought can be caused by different types of meteorological drought, e.g. overall deficitis of rain, erratic temporal and spatial pattterns, and dry spells.
We take the union of the different sources. With the exception of 2012 as we assume this refers to the drought observed in 2011. We thus get as drought years list: 1968-1973, 1983, 1984, 1993, 1997, 2001, 2004, 2005, 2009, 2011, and 2017. Note that three of the six sources only include data after 2004.
[We could further dive into the FN reports.. basically now we only checked the years for which there were CERF allocations] [what to do with the drought declarations? I am not convinced by the 2005 year.. might be caused by 2004? So far we have left it out, but not sure how to handle it ] [from the 2015 multi risk plan: Les plus grandes sécheresses des dernières années ont été enregistrées dans les années1968-1969, 1972-1974, 1983-1984, 2008-2009, et 2011-2012. Why is this not consistent with the 2017 one :cry:]
Indicator data sources
Observed end-of-season indicators
We have 7 proposed trigger indicator values: below average rainfall, late onset, long dry stretches, end of season rainfall, WRSI and Biomass. We decided to first focus on the indicators for which we know there is wide availability of forecasts, which are below average rainfall, WRSI, and Biomass. We also added NDVI as this came up during several discussions with the team. Given the definitions, WRSI, Biomass, and NDVI are expected to have some correlation.
Below average rainfall
As data source to determine observed below average rainfall, we used CHIRPS. CHIRPS is raster data at 0.05 degree resolution and is available since 1981. We compute for each of the raster cells whether it observed below average rainfall for each 3-month period (commonly referred to as season). We then aggregated this to the area of interest, expressing the percentage of the area that observed below average rainfall. The climatological reference period to compute the terciles was 1982-2010. The graph below shows the percentage of the area of interest experiencing below average precipitation for the JJA and JAS season. The bars colored red are those indicated as drought years from our previous exploration.
From these graphs, we can conclude several things:
- In the 80s there was a particularly dry period. This is also confirmed by other sources.
- Since 2000 the 5 years with the largest percentage of below average rainfall were 2000, 2004, 2008, 2013, and 2015.
- Since 2017 we haven’t seen any significant percentage of below average rainfall.
- There is a limited correlation between below average precipitation and the list of socio-economic droughts. It is important to note here that the list of historical droughts is not fully validated and thus not too many conclusions should be drawn from it. For example the 2001 drought might refer to the 2000 rainy season. Nevertheless, for example in 2017 we know there was localized meteorological drought that is clearly not reflected in the below average seasonal precipitation over the area of interest.
Biomasse
Biomasse is based on a productivity measure of dry matter production (DMP) over 36 dekads in the year, measuring average daily production in each dekad (from VEGETATION sensor of PROBA-V). We use the biomasse anomaly which is the cumulative biomasse until that dekad for that season, divided by the mean cumulative biomasse up until that dekad across all years. Seasons are defined as starting in the first dekad of April.
Biomasse is specifically designed with pastoralist areas in mind and measures more specifically the impact to vegetation than rainfall does. The Biomasse data is made available near the end of the following dekad at ADMIN2 level, meaning a publication delay of about 20 days. We download this data and aggregate it to the area of interest by taking the sum across the administrative areas of interest.
As we can see from the graph above:
- The anomaly significantly differs per year;
- In several seasons we see a dip earlier in the season but this recovers later, i.e. begin of season values are not always representative for the end of season state;
- An anomaly of <= 40 as proposed by the working group is never reached
We can also compare the Biomasse values in November to the list of socio-economic droughts. Since the anomaly of 40 is never seen, we changed it to 80 as this gave the best correspondence with the historical socio-economic droughts. The confusion matrix is shown below. From this we can see that there is a decent correlation.
WRSI
The WRSI is based on the use of precipitation and potential evaporanspiration (PET) data. Previously, these were sourced from NOAA CPC rainfall estimates (RFE) and Global Data Assimilation System (GDAS) PET data, however now precipitation data is sourced from CHIRPS and NOAA reference evapotranspiration (ETO) data.
The data is made available near the beginning of the following dekad, with only a 1 or 2 day delay. Data is generated separate for cropland and rangeland, with both parts overlapping. For the purposes of comparison, we use the WRSI anomaly, which is the % of the extended WRSI in the current dekad relative to the median in all other years.
NDVI
Comparison of the data sources
Now that we have analyzed the 4 indicators individually, we can see how they correlate. If two or more indicators show a same pattern, then we prefer to simplify the trigger by not including all of them.
We can firstly do a very simple comparison by looking at the 5 worst years for each indicator since 2000, see below. We can see that none of the indicators match perfectly with the list of drought years. However, 3 out of 5 years were picked up by Biomasse as well as NDVI (2004, 2009,2011). The other two drought years (2001 and 2017) weren’t the worst years for any of the indicators. For 2017 this might be explained by the notion of very localized drought. For 2001 we don’t know the cause of the drought and whether it was caused by the rainy season of 2001 or 2000, so it is hard to pinpoint why this is not identified as meteorological drought by the different indicators.
We can see that especially below average rainfall shows little similarity to the list of drought years and also less to the other indicators. [Is there a better way to visualize this correspondence? E.g. a gigantic confusion matrix? ]
It is not only about the match of years, but also about the separability between drought and non-drought years that shows the quality of the indicator [add some of Seth’s green/red graphs? ]
Forecasted indicators
Note: naming it forecast is bit misleading
Below average precipitation
In the previous section we saw that the observed below average precipitation in JJA and JAS doesn’t correspond very well with our list of historical socio-economic droughts. Nevertheless, it was decided to inspect the use of forecast of below average precipitation as this is the only widely-used product that is available well-ahead of time. Several seasonal forecast providers exist and the proposed provider was IRI. This data is only available since 2017.
Since 2017 these forecast didn’t predict any high probability of below average precipitation during JJA and JAS. This is in line with our observations, which didn’t see any observed below average precipitation since 2017. However, due to this limited data we don’t have any events we would have wanted to activate. Therefore, it is hard to judge the quality of the forecast for our specific use case and even harder to set an appropriate threshold.
We therefore move to literature to base our findings on. One thing we can look at is a measure of skill, the GROC, which is provided by IRI as a map. See below this GROC for JJA and JAS with a leadtime of 2 months.
We would like a GROC of at least 0.5, so none-grey. We can see that in both cases a part of Chad has a GROC that is larger than 0.5, but also some parts that are grey. Especially during JJA this area is larger. These scores also differ per leadtime, where the shown leadtime of 3 months is one of the leadtimes having a relatively high GROC. Purely looking at these GROC’s we can conclude that the forecasts can be used, but they should be used with caution, especially the one for JJA.
One disadvantage of the GROC is that it groups all three tercile categories: below normal, normal, and above normal. We are only interested in the below normal category and thus it would be good to know more about the skill for only that category. IRI only provides this across all seasons, for which the skill for below normal is even worse. However, there can be a large difference in predictability between seasons.
Another source we can retrieve information from is a paper by the UK Met Office titled Assessing the Skill and Reliability of Seasonal Climate Forecasts in Sahelian West Africa. They assess the skill of several seasonal forecasts covering Chad. They only look at the JAS season with a leadtime of 2 or 3 months. In this paper the IRI model is not included, but one model that is a part of the IRI model is included. After discussion with both the UK Met Office and IRI not large differences are expected between the models and we can thus transfer these results with caution.
The paper shows that the ROC of the below normal tercile is higher than that of the normal tercile, which is positive for our purpose.
Lastly, it should be noted that there is a general notion that the forecast skill might be influenced by ENSO and Sea Surface Temperatures (SST) in other areas. There is a general notion that during El Nino / La Nina states the forecast is more reliable than during Neutral states. This was noted by several climate scientists, while we didn’t find a paper stating this clearly. For this trigger we left it out of the scope as it would complicate things too much, but it is an important notion and should be investigated further.
TODO: write about PRESASS
Decisions to be made: - Do we include PRESASS? - Do we include JJA and/or JAS? - Do we evaluate the trigger at several points in time? - How much money do we attach?